Contextual chat with ChatGPT / GPT-3 of OpenAI API.

NOTE:

• This feature is an experimental version.
• Parameters for GPT-3 Completion API will be ignored.
• We may change the interface in the (near) future.

Install.

$ pip install gpt3-contextual

Make script as console.py.

import asyncio
from gpt3contextual import ContextualChat, ContextManager

async def main():
    cm = ContextManager()
    cc = ContextualChat("YOUR_OPENAI_APIKEY", context_manager=cm)

    while True:
        text = input("Human> ")
        resp, prompt, completion = await cc.chat("user1234567890", text)
        print(f"AI> {resp}")

asyncio.run(main())

Run console.py and chat with GPT-3.

human> hello
AI>  Hi, how can I help you?
human> I'm hungry
AI>  What would you like to eat?
human> sandwitches
AI>  What kind of sandwich would you like?
human> ham&egg        
AI>  Would you like that on a white or wheat bread?
human> wheet
AI>  Would you like anything else with your ham and egg sandwich on wheat bread?
human> Everything is fine, thank you.
AI>  Great! Your order has been placed. Enjoy your meal!

If you want to use ChatGPT, use ContextualChatGTP instead of ContextualChat like below:

import asyncio
from gpt3contextual import ContextualChatGPT, ContextManager

async def main():
    cm = ContextManager()
    cc = ContextualChatGPT("YOUR_OPENAI_APIKEY", context_manager=cm)

    while True:
        text = input("Human> ")
        resp, prompt, completion = await cc.chat("user1234567890", text)
        print(f"AI> {resp}")

asyncio.run(main())

You can set parameters to customize conversation senario when you make the instance of ContextManager.

• timeout: int : Stored context will be ignored when timeout(sec) passed after last updated.
• username: str : Name of roll of user. Default=Human.
• agentname: str = Name or role of agent(bot). Default=AI.
• chat_description: str : Some conditions to be considered in the senario of conversation.
• history_count: int: History count to use in prompt.

If you want to change these values for specific user(context) at runtime, call ContextManager#reset.

context_manager.reset("user1234567890", username="兄", agentname="妹", chat_description="仲のいい兄と妹の会話です。丁寧語は使いません。")

And, you can customize OpenAI specs to pass the parameters to ContextualChat. See also https://platform.openai.com/docs/api-reference/completions to understand more.

• api_key: str : API key for OpenAI API.
• connection_str: str : SQLAlchemy connection string for database. Default=sqlite:https:///gpt3contextual.db.
• context_manager: ContextManager : ContextManager you build.
• model: str : The AI model to use. Default=text-davinci-003.
• temperature: float : What sampling temperature to use, between 0 and 2. Default=0.5.
• max_tokens: int : The maximum number of tokens to generate in the completion. Default=2000.
• **completion_params: Other parameters for completions if you want to set.

Especially, to use other RDBMS instead of SQLite, set connection_str like bellow:

cc = ContextualChat(
    "YOUR_OPENAI_APIKEY",
    connection_str="postgresql:https://user:password@host:port/dbname",
    context_manager=cm
)

GPT-3 has capability of various kinds of task such as chat, research, translation, calculation, games and so on. You can switch the "mode" by setting username, agentname and chat_description like below.

# Translation
cm = ContextManager(username="English", agentname="Japanese", chat_description="Translate from English to Japanese.")
# Calculation
cm = ContextManager(username="Question", agentname="Answer", chat_description="Calculate.")
# Customizing end of sentence (Japanese)
cm = ContextManager(username="兄", agentname="妹", chat_description="これは兄と妹との会話です。妹は語尾に「ニャ」をつけて話します。")

Set your OpenAI API key and secret access key for configuration endpoints to example/api.py

# Settings
openai_apikey = "SET_YOUR_OPENAI_API_KEY"
config_access_key = "CHANGE_THIS_VALUE_AS_YOU_LIKE"

Install dependencies.

$ pip install uvicorn fastapi

Run.

$ uvicorn run:app

See API document and try APIs. 👉 https://127.0.0.1:8000/docs

If you want to change IP address or port, start uvicorn like this:

$ uvicorn run:app --host 0.0.0.0 --port 80

And, if you want to run this API in multi-process, use database for context management. See gpt3contextual.context.SQLiteContextManager and customize it for the RDBMS you want to use.

Enjoy your own contextual GPT-3 API👍

This chapter is based on text-davinci-003

As you know, Completion endpoint doesn't provide the features for keep and use context. So, simply, this library send the previous conversation histories with the request text like below.

Prompt on the 1st turn:

human:hello
AI:

Chatbot returns Hi, how can I help you?.

Prompt on the 2nd turn:

human:hello
AI:Hi, how can I help you?
human:I'm hungry
AI:

Chatbot returns What would you like to eat?

:

Prompt on the 6th turn:

human:sandwitches
AI:What kind of sandwich would you like?
human:ham&egg        
AI:Would you like that on a white or wheat bread?
human:wheet
AI:Would you like anything else with your ham and egg sandwich on wheat bread?
human:Everything is fine, thank you.
AI:

Chatbot returns Great! Your order has been placed. Enjoy your meal!

If you change the username and agentname the senario of conversation changes. And, setting chat_description also make effects on the situations.

Here is an example to simulate the conversation by brother and sister in Japanese.

cm = ContextManager(
    username="兄",
    agentname="妹",
    chat_description="これは兄と親しい妹との会話です。仲良しなので丁寧語を使わずに話してください。"
)
cc = ContextualChat("YOUR_OPENAI_APIKEY", context_manager=cm)

Prompt to be sent to OpenAI API is like bellow. chat_description is always set to the first line, no matter how many turns of conversation proceed.

これは兄と親しい妹との会話です。仲良しなので丁寧語を使わずに話してください。
兄:おはよー
妹:おはよー！今日の予定は？
兄:いつも通り、特にないよ
妹:じゃあ、今日は一緒に遊ぼうよ！
兄:何かしたいことある？
妹:うん、今日は映画を見よう！
兄:いいね。 どんなのがいい？
妹:思いついた！サスペンス映画がいいかな！